LLM-DSE: Searching Accelerator Parameters with LLM Agents

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Wang, Hanyu, Wu, Xinrui, Ding, Zijian, Zheng, Su, Wang, Chengyue, Prakriya, Neha, Nowatzki, Tony, Sun, Yizhou, Cong, Jason
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866909915779432448
author Wang, Hanyu
Wu, Xinrui
Ding, Zijian
Zheng, Su
Wang, Chengyue
Prakriya, Neha
Nowatzki, Tony
Sun, Yizhou
Cong, Jason
author_facet Wang, Hanyu
Wu, Xinrui
Ding, Zijian
Zheng, Su
Wang, Chengyue
Prakriya, Neha
Nowatzki, Tony
Sun, Yizhou
Cong, Jason
contents Even though high-level synthesis (HLS) tools mitigate the challenges of programming domain-specific accelerators (DSAs) by raising the abstraction level, optimizing hardware directive parameters remains a significant hurdle. Existing heuristic and learning-based methods struggle with adaptability and sample efficiency. We present LLM-DSE, a multi-agent framework designed specifically for optimizing HLS directives. Combining LLM with design space exploration (DSE), our explorer coordinates four agents: Router, Specialists, Arbitrator, and Critic. These multi-agent components interact with various tools to accelerate the optimization process. LLM-DSE leverages essential domain knowledge to identify efficient parameter combinations while maintaining adaptability through verbal learning from online interactions. Evaluations on the HLSyn dataset demonstrate that LLM-DSE achieves substantial $2.55\times$ performance gains over state-of-the-art methods, uncovering novel designs while reducing runtime. Ablation studies validate the effectiveness and necessity of the proposed agent interactions. Our code is open-sourced here: https://github.com/Nozidoali/LLM-DSE.
format Preprint
id arxiv_https___arxiv_org_abs_2505_12188
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle LLM-DSE: Searching Accelerator Parameters with LLM Agents
Wang, Hanyu
Wu, Xinrui
Ding, Zijian
Zheng, Su
Wang, Chengyue
Prakriya, Neha
Nowatzki, Tony
Sun, Yizhou
Cong, Jason
Hardware Architecture
Artificial Intelligence
Even though high-level synthesis (HLS) tools mitigate the challenges of programming domain-specific accelerators (DSAs) by raising the abstraction level, optimizing hardware directive parameters remains a significant hurdle. Existing heuristic and learning-based methods struggle with adaptability and sample efficiency. We present LLM-DSE, a multi-agent framework designed specifically for optimizing HLS directives. Combining LLM with design space exploration (DSE), our explorer coordinates four agents: Router, Specialists, Arbitrator, and Critic. These multi-agent components interact with various tools to accelerate the optimization process. LLM-DSE leverages essential domain knowledge to identify efficient parameter combinations while maintaining adaptability through verbal learning from online interactions. Evaluations on the HLSyn dataset demonstrate that LLM-DSE achieves substantial $2.55\times$ performance gains over state-of-the-art methods, uncovering novel designs while reducing runtime. Ablation studies validate the effectiveness and necessity of the proposed agent interactions. Our code is open-sourced here: https://github.com/Nozidoali/LLM-DSE.
title LLM-DSE: Searching Accelerator Parameters with LLM Agents
topic Hardware Architecture
Artificial Intelligence
url https://arxiv.org/abs/2505.12188